Found first: a primary source the press has not covered yet.
Researchers at the National University of Singapore built a three-layer harness that makes force sensing available to frozen frontier models as a standard tool call, with no fine-tuning or demonstrations required. On a real robot arm, it raised GPT-6 Astra's zero-shot contact-rich task success rate from 3.3% to 90.0%. The paper is arXiv:2611.21099, submitted November 16, 2026.
What the source says
Ke Sun, Kevin Yuchen Ma, Yike Zhang, Yanzhe Chen, and Mike Zheng Shou designed ForceHarness with three components: self-terminating force primitives that run fast force-control loops below the API's decision rate, a physics translation layer that converts raw force traces into physical quantities with uncertainty estimates, and a tool-call result that returns those quantities to the frozen model. On HiddenPhys, a seven-task MuJoCo simulation benchmark, it brought performance from 62.1% to 96.4%, matching hand-written expert force controllers and surpassing Diffusion Policy, π0.5, and ForceVLA each trained on roughly 200 demonstrations per task. On the public ForceBench suite, it added 17.5 percentage points. Ablation results show smaller models benefit more from receiving physical quantities than from raw force values.
Why it matters
Contact-rich manipulation has been a consistent failure mode for API-only frontier models because force data was unavailable above the hardware abstraction layer. ForceHarness shows the gap closes when the tool-call interface is redesigned, not the model. The jump from 3.3% to 90.0% zero-shot, with no demonstrations, is directly relevant to teams deploying frontier models on physical hardware through an API. The pattern that smaller models improve more from physical quantities than from raw traces points to a representational bottleneck, not a scale one.